Abstract
Masson staining is crucial for accurately identifying fibrotic areas in kidney tissue. The CycleGan model can convert HE-stained kidney pathology images into Masson-stained kidney section images, aiding doctors in identifying fibrotic regions. Prior models had limited generalization performance, causing inaccuracies in simulated images when there was a color shift between training and validation sets. This paper introduces SPCGan, a CycleGan-based staining translation model. SPCGan incorporates a color enhancement algorithm during training to increase the number of HE-stained kidney pathology image samples. Additionally, the normalization layer in CycleGan is replaced with the Spade layer to prevent the loss of spatial morphology semantic information when image features pass through the normalization layer. After validation on a non-public dataset, SPCGan improved staining accuracy for StainGan by 0.034 and 0.051 for CNGan, with FID scores improving by 1.47 and 10.26, respectively. This enhanced generalization performance and improved the similarity between simulated Masson staining results and authentic images, making significant progress in histopathological image staining translation.